Building with AI · Agent UX patterns

Tool-call surfacing and undo

You can name the four UX layers that a trustworthy agent surfaces (intentions, actions, results, recovery) and explain why undo is a feature of action class, not a feature of UI.

Once you accept plan mode, the next question is: during execution, what does the user see? A 10-minute autonomous run is opaque by default. Without a surfacing strategy, the user is stuck choosing between staring at a spinner and walking away with no idea whether to trust the result.

Anthropic's Trustworthy Agents in Practice proposes a four-layer model that maps cleanly onto UX:

- Intentions — what the agent is trying to do, right now and next. - Actions — what tool calls it is making, with arguments. - Results — what came back from each call. - Recovery — how the agent reacts when things go wrong, and what the user can do to step in.

Real products in 2026 implement this in different ways. Claude Code's Agent View renders a chronological activity stream with the four layers visible at different zoom levels. Cursor 3 moves the action layer into an agent workspace with local/cloud agents, diffs, screenshots, and PR handoff. Notion and Linear render agent activity in their existing notification feeds. The patterns are converging.

Chapter contains 4 lessons.